DeepSeek's 3T Model: The Scale-Up That Has to Wait for the Clusters

On September 14, Deep​Seek will retire DeepSeek V4 Pro — the 1.6-trillion-parameter flagship it put into general availability on August 13 — and answer every call to deepseek-v4-pro with DeepSeek V4.1 Flash, the 552-billion-parameter model it released on September 10. Four days before that switch was announced, a pseudonymous Deep​Seek watcher posted something pointed in the opposite direction: that the lab had been working on a 3-trillion-parameter model, "probably another 1T params Engram," and that the reason it is not out is economics rather than capability.

Nothing about that model is confirmed. There is no name, no repository, no config file, no benchmark, no release window — one X post, sourced to "good authority," with its single most interesting detail explicitly flagged by its own author as speculation. Treat it as a signal, not a fact. It is worth reading anyway, because the two halves of the claim pull in opposite directions and only one of them is likely to survive contact with Deep​Seek's own roadmap.

What the leak actually says, and what it does not

The post comes from the account teortaxesTex, a long-running Deep​Seek watcher who describes himself as a fan of the lab since 2023. It reads in full: "Actually I have it on good authority that Deep​Seek was working on a 3T model (backbone, probably another 1T params Engram but that's my speculation). They just weren't sure it'd be economical to post-train and serve it. When their clusters grow, we'll see some..."

Four things in those two sentences carry weight, and one of them is not a fact at all. "Good authority" is unnamed. The parameter count arrives as a single number with no split between total and active. The Engram aside is labelled speculation by the person doing the speculating. And "when their clusters grow" is a conditional with no date attached to it.

• What is being claimed — that a 3-trillion-parameter Deep​Seek model exists in some stage of development.

• What is explicitly the author's own guess — that roughly 1T of it is Engram memory.

• What is unknown — its name, architecture, active-parameter count, context window, training status, and whether it ships as a product at all.

• What is verifiable right now — nothing. No Deep​Seek repository, model card, pricing row or docs entry mentions it.

Even the arithmetic is ambiguous. "3T model (backbone, probably another 1T params Engram)" supports two readings: a 3T model of which about 1T is Engram, or a 3T backbone with another 1T of Engram alongside it, for roughly 4T in total. That gap is a trillion parameters wide, and it changes the serving bill by more than most labs spend on a training run.

It is also worth remembering that this account's record is mixed rather than oracle-grade. He read Deep​Seek's September 9 notice about rerouting V4 Pro traffic as evidence the lab had "solved architectural issues of the V4 family" — a reasonable inference, and still an inference. In early September he also floated the idea that an anonymous stealth model on a third-party coding platform was Xiaomi's MiMo-V3-Flash, which nobody has confirmed. Useful early signal; not a source of record.

The interesting half is the memory table, not the parameter count

Engram is real, and it is the reason a 3T claim is more plausible than it first sounds. Deep​Seek published the conditional-memory architecture in January 2026 with open-source code attached, and it does something unusual: it separates static knowledge retrieval from dynamic reasoning. Instead of spending GPU compute recalling facts, the model hashes its context into embedding tables held in DRAM and retrieves in constant time, with no GPU work in the lookup path, plus a gating mechanism that suppresses a retrieved memory when it conflicts with the surrounding context.

The figures Deep​Seek reported for its own test configuration are the ones to hold on to: a 100-billion-parameter embedding table offloaded to DRAM at under 3% throughput penalty, and 97% needle-in-a-haystack accuracy at 1M tokens against 84.2% for standard attention. Those are the lab's own numbers, not reproduced by us. Independent early evaluations have reportedly landed in the 93–96% range at the same context length — clear of baseline, below the claim.

Scale that idea up by ten and the shape of the leak changes. If most of a 3T model's extra parameters are hash-table memory sitting in DRAM rather than experts competing for HBM, then "3T" stops being a proxy for "unservable." Total parameter count and serving cost come apart. That is the genuinely new idea in this leak, and it is the part the published research actually supports.

The caveat is that the Engram paper reportedly does not address inference-time overhead — latency and throughput — which is precisely where a DRAM-resident lookup table would show up if it showed up anywhere. Memory you have to go and fetch is not memory you get for free.

Why Deep​Seek's own September argues the other way

The case against a 3T product shipping soon is that Deep​Seek just ran the experiment at 1.6T and answered its own question with something smaller. The V4 ladder as it stands today:

• DeepSeek-V4-Flash-0731 — 284B total parameters, 13B active per token.

• DeepSeek-V4-Pro-0813 — 1.6T total, 49B active, open weights under an MIT licence.

• DeepSeek V4.1 Flash — 552B backbone on a Causal Encoder-Decoder design that activates 8B parameters per token during prefill and 16B during decode.

On September 14 at noon Beijing time, the middle rung comes out. Deep​Seek takes V4 Pro offline and routes requests for deepseek-v4-pro to V4.1 Flash, billed at Flash rates, until a V4.1 Pro exists. The official pricing page today carries two rows, and neither is the retired model's successor by size: deepseek-flash at $0.30 per million input tokens and $1.20 per million output at peak, deepseek-v4-pro at $1.32 and $3.96, with off-peak rates at half those figures. Both list a 1M-token context window and a 384,000-token output ceiling.

The direction of travel is not subtle. A lab that retires its largest model in favour of one activating a tenth as many parameters per token has already concluded that the economical unit of frontier capability is not the biggest checkpoint it can train. A 3T model whose builder is unsure it can be "post-trained and served economically" is not a rumour about hesitation; it describes a lab that now has measured data on the alternative.

Post-training is the sharper half of that problem. Full-parameter post-training of the DeepSeek-V4-Pro line on Huawei's CloudMatrix384 SuperPOD has been reported by the third-party SLAI T-Rex project at 34.22% MFU, roughly 2.93x the open baseline recipe. That is an encouraging number — a third-party number, on a 1.6T model. Doubling the backbone roughly doubles the bill before a single token is served.

"When their clusters grow" is the whole sentence

The load-bearing clause in the leak is the last one, because it is the only part with a public paper trail. Deep​Seek is reported to be planning at least 160,000 of Huawei's Ascend 950DT accelerators at a new data centre in Ulanqab, Inner Mongolia, in an order valued at roughly $2.56 billion — one of the largest clusters of Chinese-made AI silicon anyone has attempted.

The caveats attached to that plan matter more than the headline. The chips are planned for inference, not training: Deep​Seek has tried training on Huawei silicon before and still trains on Nvidia accelerators, which is the step a 3T model would actually need. Delivery could take more than a year. Partial capacity is expected online in late 2027 into early 2028. And high-bandwidth memory supply, not logic, is expected to cap how many 950DT units Huawei can build this year.

So the optimistic reading of "when their clusters grow" puts a 3T model on a 2027–2028 clock at the earliest, and the pessimistic reading — that it remains a research artefact and never becomes a product — is consistent with everything Deep​Seek has shipped in 2026. The compute environment around the lab is also contested policy rather than a pure supply question: on September 8–9 the NSA, FBI and CISA jointly alleged that six Chinese labs including Deep​Seek distilled US frontier models at "industrial scale," an accusation China's commerce ministry rejected. Whatever one makes of the allegation, a release that depends on cluster growth depends on decisions nobody inside Deep​Seek controls.

What would turn this from a leak into a launch

The checklist is short and specific, and none of it has fired yet:

• A repository under the deepseek-ai organisation on Hugging Face, with a versioned checkpoint name rather than a family slug.

• A new row on Deep​Seek's Models & Pricing page.

• A dated entry on the API docs news feed, in the lab's usual newsYYMMDD format.

• A model identifier, not a family name — Deep​Seek versions checkpoints, and a bare family slug has so far meant a preview.

The nearer milestone is V4.1 Pro, which a Deep​Seek team member referenced on September 9 as the endpoint of the routing window. It has no published specs, no parameter count, no price and no date either. In one sense V4.1 Pro is the test of this leak: if the next flagship lands at roughly the V4 Pro's 1.6T or below, the 3T claim has slipped at least a generation.

What any of this means if you are choosing a model this week

A 3T model is not a purchasing decision in 2026, and the practical lesson of Deep​Seek's September is not about scale at all. It is that the model behind an endpoint can change while the endpoint's name stays exactly the same. Anyone calling deepseek-v4-pro through an abstraction layer gets a different, smaller, cheaper model on September 14 without touching their code. Anyone wired directly to a single provider's implementation of that ID gets whatever that provider decides to do.

Both DeepSeek V4.1 Flash and DeepSeek V4 Pro are on OrcaRouter under one key at 0% markup — provider list price passed through, which means Deep​Seek's September 10 pricing change is live here the same day at list rather than at some marked-up version of it. The model page reads $0.15 per million input tokens and $0.60 per million output in the off-peak band, which is Deep​Seek's own off-peak number and nothing on top. Automatic failover across providers is the unglamorous feature that matters most in a week like this one: when a vendor swaps the model out from under an endpoint, the routing layer is where that becomes a config change instead of a migration.

If a 3T Deep​Seek model ever does ship, it will be served by whoever has the clusters to hold it, at a price the compute dictates. The same routing layer is where you would meet it — without a second contract and without rebuilding anything.

The open question is not whether Deep​Seek can build a 3-trillion-parameter model. Three published architecture papers, a working memory design and a 552B model that its maker says outperforms its own 1.6T predecessor all suggest the lab knows how. The question is whether anyone will pay to serve it — and on the evidence of the last two weeks, Deep​Seek itself is not sure. Watch the cluster, not the parameter count. The pricing page will tell you what Deep​Seek actually ships faster than any leak will.